.Net Foundation Design Reviews, 26 Nov 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: .NET Design Review: GitHub Quick Reviews

Description

00:00:00 - Approved: Add asynchronous overload of WindowsIdentity.RunImpersonated https://github.com/dotnet/corefx/issues/24977#issuecomment-558756052
00:12:30 - Approved: Consider Socket.OSSupportsUnixDomainSockets property https://github.com/dotnet/corefx/issues/27643#issuecomment-558760298
00:23:36 - Approved: System.Threading.Channels : Expose Channel.EstimatedCount https://github.com/dotnet/corefx/issues/30852#issuecomment-558762980
01:01:14 - Approved: Support Rfc3279 signature format for DSA and EcDSA https://github.com/dotnet/corefx/issues/42684#issuecomment-558782564
01:21:06 - Approved: Consider exposing a HWIntrinsic that allows efficient loading, regardless of encoding. https://github.com/dotnet/corefx/issues/33566#issuecomment-558793376

A

A

A

Confusion inspired by the fact that we used to have some bugs around this I. Don't know if he's still, hopefully we don't still, but we at least used to have some bad ones that have been fixed where.

A

While you're impersonating, you know if this function kicks off asynchronous work, then that work will also when it runs the execution context will flow with it and so it'll also be running in the impersonated context, which is what we want.

A

What that means is you can actually pass an async lambda as the func here, and it will work exactly the way you want it to work. So you can wait await this thing.

A

You know say a wait: Windows identity, don't run person ate it as an async lambda, and because of then how all this composes, the when you have an async lambda, that's gonna, be a func of task of T, so run impersonated is going to return a task of T, which you then await, and you know the results of the wait is the T and the whole body of the await ends up correctly of the the async method ends up correctly being impersonated, because again it just sort of naturally falls out.

A

You impersonate you, the continuation gets kicked off as part of that impersonation and so that continuation will run while it's impersonated, so you can use the existing API.

A

Fine, a concern is it's not at all obvious that that works, and so the proposal that was made by someone else and I'm just clarifying their intent, and my comment here was expose an async API that returns tasks strongly typed, and even if the implementation, literally just delegates to the existing method, it'll be more obvious. What it's doing.

B

But basically you're not opposed to adding them you're just saying it's relatively cheap to implement them. I'm.

A

Saying technically it's unnecessary: it's not adding any it's, not it's! Adding zero functionality over what's already been added! Oh it's already there, but I agree that what's there I'd be viewed as no people might not understand what it's doing.

C

And so I don't have.

A

A I don't think it's bad to add these I, don't think it's bad.

D

But it seems that this is a feature in general for somebody who already is searching for this. Oh no, what they are doing I mean it's not something where we like worry about. Discoverability of just you know, looking around and stamping up.

A

Yes, and no, there is another API that at least Barry and I can't remember as well. I think leave as there is. You can speak to it, um we're against there's the Windows identity, dot impersonate, which we did not bring over from done in framework for security concerns, I, believe and so developers that we're using Windows identity, dot impersonate in in an ASIC method, now go searching for an alternative and they only find a run.

A

Impersonated method that looks like it takes a synchronous call back when what they wanted was asynchronous, and so then they get stuck Steve.

E

Can you remind me so this method run impersonate I'm, sorry, the original method run impersonated the text funk of teeth that also exists on full framework, starting in four six. What happens if I pass an asynchronous method in there? It's.

A

I can double check, but I'm 99% sure that it also works just fine, because it's just using you know not in framework also flows, identity as part of execution context. Impersonation is kind of execution context. We just do it in a different way in core in desktop or in done. If framework there's a lot of complicated additional stuff as part of execution context of lo it in core, we basically just shove it into an async local and then allow the async local machinery to propagate, because.

E

The argument I would have made is if it behaves differently in full framework vs. core than these methods could be because Ariel a compilation, failure all right. Actually you put into the providers. Oh you would. You would get a couple extra player if you reach target is like that standard. So.

A

Yeah, um no, you know if, if we choose to expose these, then that raises the question well what about, for example, cancellation, tokens but I'm, not sure what that would mean in these cases, I think it would just sort of be a way to avoid a closure allocation, because generally someone's gonna be using this with a lambda, and so they would just close over whatever cancellation token they wanted to use and.

F

A

Only reason we would then want to thread it through would be to avoid the closure and delegate allocation, but.

D

If you're talking.

A

About impersonation, it's pretty expensive, already I.

D

Mean so I would say another reason for having a cancellation talk and explicitly is that it communicates to people how they should properly. You know, use the API properly. You think you know it basically just says it's a good practice to pass cancellation tokens whatever you do, potentially.

G

D

G

A cancellation.

G

A

A funk of cancellation token, two tasks of T, func, comma cancellation, token I'm sort of I could go either way. Yeah.

E

The problem with that is again what you said about like over here. We can just pass the cancellation token.

H

But we just say we don't leave this at all. It just works and every time anybody opens the issue asking for one of these it works.

H

D

What do you mean by it use the existing method? Oh well, something given what's Steve said about existing API into that ephemeral, but now it's gone I actually find the argument that it would be useful to just add those. If it was like you know, we never had this API I would almost say you shouldn't add it, but, as he said, the framework doesn't have any personnel yeah.

D

A

I

A

Api you, you can use in.

A

F

E

Yeah, okay yeah it does, it does improve discoverability, and that is a good boy.

E

E

But we decided a or nay on the campus since we're already capturing the correlated state. How.

H

Would you even like Eve if we took the casual Asian token as a parameter here, I guess, if the funk, so there was a cancellation token yeah.

A

The concern is, it adds a lot of ceremony because you know you you know, rather than just doing you know, delegate, open, brace or acing or lambda. You know open paren, close paren. Now you say token and then you have at the end comma cancellation token.

A

So I don't know, and it's not like we're if, if this method were going to be doing additional work, that was itself cancelable. That would be a good argument for taking the cancellation token, because then we would use it ourselves as part of the implementation of this method, but I don't believe, there's a situation in which we would the best we would do is you know, throw if the cancellation token was already cancelled.

J

Right there again, everybody, you would add the destiny. Looking back in the future, I'm so worried there.

I

Would be being two more.

A

K

E

If we did that, we would probably also have state object and all those who avoid that voters are about look you're doing an impersonation you're already making at least phone calls.

D

If you'll design guidelines, don't Christopher's consolation to hold the tiny legs and I.

L

A

This is an interesting one, because there's you know the the the thing that is a sink here is provided by the collar.

H

This is a sink callback which, like the only thing you had in the guidelines, for it, is, if you're converting a thing that took a synchronous, call back making you're, making an asynchronous version of that. Consider, accepting an asynchronous call back.

H

D

Like unveiling.

A

The also I guess it's I think it's reasonable to do this now and then you know if we get feedback in the next year, we could consider adding anyway.

D

It's very similar to the async event pattern that it's literally almost the same, except it's a method that makes a function as opposed to property. This concept also fine.

D

A

This also execute sit like promptly as part of passing it, whereas the event one you were talking about is a callback at some later point in time.

K

B

Does anybody else use my friendless building.

B

Last four days, Wi-Fi clips in and out.

B

Last week or two the call dropped because the wife had collapsed. So I just got a message from OBS that it's reconnecting. So it's not encouraging confidence.

B

M

Working from home this weekend.

N

I'm gonna go grab one of the internet, but after hide.

B

You're so cute, this is a Mick bull. You can't just plug it in you're, okay,.

B

There's an adapter you can plug in, but of course you have to be boo before the credit before the deaf diversity can just plug this guy. In what do you think this is the 90s? No.

E

No, a friend with a nine news, you'd plug in an adapter on your machine with my screen. Only.

B

When you're plugging the scanner all right, so then, let's close this going here, you seems your app again.

A

All minutes from long ago.

B

A

I, honestly, don't have a strong opinion about this. One either is just that we have these existing properties on socket OS Sports ipv4 OS supports ipv6, someone I, don't remember, who made a request on some PR somewhere that we also have one for OS supports UNIX domain sockets Windows now supports UNIX domain sockets as of Windows about a year ago, Windows 10 about a year ago. So this would return false on older versions of Windows said.

B

To me, he's already reason enough.

F

This one seems like a fairly.

E

This remembering the way Lele back when, when windows used to have like Windows 98 used to have support for almost unit domain sockets, and then they ripped it out. Because now when was they.

B

Like almost that's, that's Glamis, we have any care for this almost supposed to mean.

A

Yes, which is almost certainly with the implementation, is going to do that's what we currently do. We just do it once in cash. It.

B

Well, Jeremy, you seem to have some objection because you were not super keen on anyone with you rationale.

O

B

Said that, but you said this after, like.

H

So the only hesitation is following up on me: there's a bunch of other capability api's that we could add, so we want to do. One now versus please be like here are the 17. We want to add well, I, think.

D

The limit is there 17, because there's tons we may want to think about other API.

G

Then just add your limb. Well, this this hell's I mean this. Does nice enough resolve it.

D

I know so, if it's just you know clearer than the for granted, yes, but if we truly think there are 17 coming in the future, then I would you know, say: I know adding 17 billion properties do.

E

We actually have a TRS that allow you to program again. We did again the.

F

A determine not to one. There are many address vendors, but that's they're.

A

K

B

When I think Germans point I think like if you just wanted to make a big border right, if you imagine you walk the list and you're like okay, if you would have to expose all of them, how would we do that? Maybe we end up and flagging them or something right versus well, we had one. We had one, we had one and then you know to leases later. It's like! Oh no. We have filled your mess because we've never seen a full list. Right I mean that would be the English. That's.

A

Totally valid one caveat to that is the domain of the the design of sockets today, with the whole endpoint kind of class hierarchy generally means that each of these different kinds of you know sockets that you're creating is gonna require a different, concrete, endpoint class like we have. You know: IP endpoint, we added UNIX domain, socket end point I. Imagine if we're gonna add another OS or it's Bluetooth, sockets or whatever, there's probably a Bluetooth endpoint that we then need to go and add, or something similar.

A

So there's probably I'm waving my hands a little bit, but there's probably going to be additional API that we end up needing to add for each of those seventeen.

H

We have this one, then we are at feature parity between the address families that we have active support for, and the capability like we.

A

K

A

Think we only have three concrete endpoint types today: DNS endpoint, IP, endpoint and UNIX domain socket endpoint that DNS endpoint is just about mapping names via DNS to ipv4, 1986 and.

E

This new proposed property goes on the socket type instead of the UNIX domain, socket in point.

P

P

It's also socket another point: oh we bi together what it's not a device type of socket correct, it's if you're on track at that point. Okay, that would be very long. Okay,.

H

I mean that could mean that one of the an answer is that we make endpoint I guess the problem is, do once attic.

H

We must implement this thing. People must call me back. Okay,.

B

I think the virtual yeah I think I need to give the paper I.

H

Guess: what's that attachment yeah, so the I guess the question is: would it be better for this live on the in point, but we already have two of them here and we think that we're gonna have three and maybe have a fourth in no more than yeah. Well,.

N

I would expect that we might have a fourth near tournament. Someone wants to actively work on blue or non Windows platforms. If we do have Bluetooth support, not in box but via the Winton API is that you can out pull down these or Nick or app, but there there's likely going to be new standards coming out in well.

N

Iii doubt that we'll end up with like just four and be done.

E

Her butt is not a problem for good. That's Henry years down there.

E

Yeah, even if that's, if that's a 10-year finale through them,.

B

You don't look here because you'll retired basically.

E

K

E

K

We will have.

E

Reinvented this entire obstruction within ten years, otherwise.

H

B

The way your next project is going to be writing a yellow, boxfish. You don't know this is coming your way. You guys all laugh but I'm actually somewhat all right. So then hey objection, approving the safety on and I guess not.

B

Let me close this comment.

I

Didn't didn't didn't.

Q

I guess divas on today.

A

Gave me so: we've had several requests for diagnostic purposes to understand how many elements are currently in a given Channel.

A

So people haven't required a comma 100%, accurate count because things are changing constantly, but they want to know approximately how many items are there? We went through a whole bunch of naming. You know start. Is this start of that count and then I made provided some feedback that maybe we should move away from that and we ended up with estimated count and then I provide a feedback that I thought about it.

A

More and count is probably fine, so I think the final thing here ended up just being account property on something channel of tea or yeah.

B

I was about to ask because it's not like it's unfair more than one generic one. No.

A

We need to be on I mean we should talk about where it would live. It would either need to be on channel of tea or Channel reader of tea, depending.

R

A

What we were hoping to achieve I think mmm-hmm. If it's how many items you want to how many items you want to be able to pull from it, and you probably want it on channel reader if we wanted this sort of it, you know how many items are in this whole thing, then we wanted on the there's a a channel of TNT out. Basically, we put on that. Probably a white.

D

Nothing that is keeping the estimated prefixes okay.

A

Like on concurrent collections, it's it's a similar thing and we just used count there because it implements the you know the normal collection interfaces, and so you know- and if, if you are using this thing in a if you like, pause the world, then you take a look at it. It's gonna be correct. It's only if you're, concurrently, changing it while you're accessing it and if you're doing that, then it's the same as accessing a concurrent collection.

A

While it's changing and you have to understand at the moment you you read it and then the moment after that it could have been changed and it's just the nature of concurrency.

D

It makes sense it's basically it's precise at some point in time. That may not be the point where you are executing the next line of code right. That's also.

N

Some water, like eight guys, we added for checking statistic.

A

I mean technically, you could imagine to tear the precise that some said precise at a moment in time I mean technically, you could have a channel implementation where that wasn't true, where you know we it's it's to compute the count. You might end up with some value that was actually never there, but that'd be pretty rare and I. Don't think it's worth worrying about is.

E

Getting this count of operation.

A

For the most part,.

A

So, like the main channel that we have is found, it is a concurrent queue underneath and concurrent queue is a linked list of segments so but the segment's grow in size.

A

So you start out with one that's 128 elements and then one that's 256 elements and then 512. So it's a relatively short list generally. We do need to walk the list. If, if there's multiple segments.

K

A

A concurrent queue is actually under the covers a linked list of circular arrays.

H

S

A

List it walks the list of segments yeah, so each segment is an o1 operation, but there might be you know, order log n segments, basically.

T

A

I mean I check to someone using this in any given out on any hot path. There were you wouldn't want to do that? You'd want to use the try methods, try, pika, try, read or whatever looking.

H

At the original justification, it's they seem to want to know like you know, not as I said, not between zero and two, but between like five and ten thousand of like Duke. Do I need to like make something start eating things fast, yeah.

E

I imagine the scenario curves that could help them really page right.

A

H

Monitoring migrating yeah.

A

H

Maybe Diagnostics.

A

No, this is gonna be pulled on some, hopefully, relatively infrequent frequency do.

E

We normally discourage people from accessing things. Like concurrent dictionary, you don't want.

A

um We do primarily because it's not as cheap as one would like like concurrent dictionary count. Concurrent dictionary is implemented as kind of a partitioned set of collections effectively. Yeah looks.

E

Like it takes a walk in order to get the fact, it.

A

Takes multiple locks and take every one of those has its own right lock, and it takes all of them.

E

So it sounds like there's already precedent with Harkin Khan collections of account. Property might not necessarily be the treatment seminar, yeah.

A

We've tried I mean you know. Obviously, we've tried remember possible to make them as cheap as we reasonably can. In fact, we had him, arguably a bug that we fixed when I rewrote concurrent queue in to oh I thought it was okay to make count much more expensive in exchange for making other things much faster, and it turned out that some people were actually using it on a hot path like if count is greater than zero. Then try it.

A

You know, try DQ, so we fixed that four to one I think it was found a way to do it without sacrificing other stuff. But you know the moral. The story is, developers will do it, even though we don't want them to.

H

Looking at concurrent, you, like our Docs for four counts mentioned, if you cared about zero down, zero use is empty. Instead, we already have his empty on channel and should we add it.

A

We don't we could list I, don't think we do what.

E

Would you do with.

H

Or colony count greater than zero and correct q, then they should have been colleague is empty, which is probably a we understand when we're transitioning to.

E

And from those.

H

E

Set this one, but then you called eq which might blow up if some other thread has empty I.

K

Don't think it'd be good.

H

Okay, yeah I mean like sure again, if you're, you know, with the justification I'm just looking at the remarks- and you know like I, was wondering if we commented about that, it's login or or what just asking question.

K

B

Wish of this API live journal of T, so.

A

That's what I'm wondering looking at the api's now I'm I'm having a debate with myself, I'm kind of wondering if we should put it on both channel reader and channel writer.

B

The penalty right now is empty right. It doesn't have anything channel.

A

T just derives from channel of T comma T, because the 99% news case is, you only have one type that goes into your channel, but technically you could have a channel that was doing a transformation, which is why we support writers and readers having different types right.

A

So we could just put it on channel of T, write, T read and then it would naturally be available to channel of T. We could put it on just channel of T couldn't say you know. It only makes sense in that use case, or we could put it on both the reader and the writer.

A

Which is arguably the most flexible, though it's probably overkill,.

B

Given that I have used the API exactly zero times so far, I do not have in fact opinion I'm, just wondering if you put it on channel of tea or film lofty right, tea, read and then later on. Somebody wants it on the reader writer and then we have it in all places which would not be helpful right. At that point we might as well add right now, through we don't know, I don't be done with it.

A

Let's, but why don't we go with that and then, let's double-check, with some of the people that asked for it to make sure that that meets their needs? You.

B

Mean we do writer or yon.

A

The reader and on the writer, oh.

E

It would be a it's, not abstract. I would imagine, because we can't break these insane implementations. No.

A

It's gonna have to be virtual and we're gonna have to have some default implementation, for example returning negative one or zero.

A

The alternative would be, you know a tri-gate count which would return false. We all of our implementations would return true, but by default return, false.

D

Maintaining zero, maybe not yeah,.

H

Either return negative one or.

A

Also, an option.

H

Like if we think that we have the only implementations and we're going to fix all of them, then Parramatta for an exception that man tells somebody hey. Your channel needs to go. Add support for this new API.

D

Are we gonna have also a tester to test whether it support it? You.

P

A

I mean if that was around we were gonna go I would make it a trying method yeah. If we were concerned about that and.

E

We think we have the only implementations, basically.

A

And we certainly have the vast majority of them in.

K

A

B

According for the company, we built flexibility.

A

Into the abstraction, but for the 99.9% case you either just want an unbounded or abounded queue and that's what we have.

A

And then write a custom own or one I, don't know. I can ask.

D

I've been doing basically processing those friends and event cards, cool.

A

Yeah I mean in general, you know, usage and feedback would be great.

E

You is the other thing that I don't want to have is a property like a virtual property, not abstract, whose implementation towards an exception, because if I am for whatever reason, sub classing this node issues intellisense a and fill in all of the abstraction. Now, all of a sudden there's some core functionality which I forgot to fill in because Eid dental well, since its estimated.

G

Comprehend 1258.

D

A

Think we were going with just count, but yeah 42 sounds like the right.

B

The trick is, it had to be activated at some point. Your number, it strikes me as 1700 was at least correct at some point: zero.

H

I would worry that negative one is a clear sign. It's not working exemptions appear time. Yeah.

D

Zero means RMB, it's just it's such a weird pattern. We don't have it anywhere else, because now the usage is get the count check, whether it's not -1. Thank you. I.

B

Mean honestly, the other.

A

B

A

I think I think try I mean I, think try gate count is the right answer. Then I was sort of on the fence about it, but if we can't figure out a good default, there's an obvious good default for try gate count. Plus it means that an implementation which, for some reason, is unable to provide account, has an out.

I

Well, we have a new get package. This.

H

Is a major version? We could start introducing new, abstract honey.

B

That's all I said like yeah is.

D

That now huh well.

B

We've made breaking changes, 13 major versions, anymore.you.

H

We've never done it I'm introducing the concept that we can start taking like the uses since I know, we've added.

I

A lot of things in my.

H

Crypto that our virtual throw because and implement them and introduce abstracts. But if you're actually inside you active all right, I mean.

B

You have another option right: another option as we can. We can introduce this idea of a soft abstract right. It's a custom attribute you put on the method and you basically in source code. It becomes a breaking change. You basically.

C

B

You basically get a warning or an error that says when you're, inheriting from this type, you know over in this number, you should alright. This man.

D

B

Because the abstract will not be in avi break change, binaries will work and bind, but in source code you may see when you recompile, you get told you need. You think.

D

We can do it because imagine that I wrote the component.

B

D

Version yeah and it didn't have this abstract, now I loved it my component into process at an upgrade of some other usable library. What.

B

You're talking about, we can't make a breaking change in binary form, let's by suggesting making it an analyzer. We.

D

Can make it a crack in it?

D

B

The default implementation could either throw which is basically what what Jeremy is proposing for the crypto API is or we could go over the try pattern, but either way. Even if you go with the try pattern, it would still say: there's a virtual try account you should still overwrite this thing right, I mean, like you know, yeah so like. But how do you know that there's right now?

B

The problem is you as soon as you extend an abstract type, one of the videos of the compiler is Atilla to use the methods you have to override and are we adding more more functionality that you just have to know that you should override it's the same fallen with text reader where the default forces you to implement? What is it right, char, whatever, but you pop on over with all the other ones as well.

B

Otherwise, you should have so really slow right, so at least it works, but like strictly speaking in the editor, you should probably get the you should do something else by.

D

The way hey one more thing for money, our current.

P

A

Nice I'd, rather just have to try method.

E

And count, and you have to over okay.

B

In that case, I would say, try call seems more natural, yeah.

D

Big island, it's literally an alternative to the desperate.

D

H

A

A

That the tester doer thing, like generally my understanding of them, is you can assume that if it returns true like you could call the thing any number of times and it's gonna work I mean that's not always true, but that's how I sort of interpret it, whereas with a try method, you know I feel like it's more understood that at any given moment it might return false, which, in my opinion, is better maps to the idea of concurrency.

A

D

Depends on the implementation of the type not on the state of the object, may be I mean for.

A

Example, imagine someone decided they wanted. You know count to be very cheap, and so, if count was going to have to take a log that was currently held by someone else, they could just say you know what I can't get the count right now, because the alternative would be to block for some long period of time. Potentially, let's.

B

See we actually sing, but I get calendars action more flexible if you ever have to be even flexible.

D

Anybody would be like, would we really implement any of our channels like that I? Don't.

A

Plan to for the three concrete implementations that we have but I think I mean there are similar things that we do in other places and so I wouldn't I, wouldn't sort of dismiss it outright. For example, there are places where there's a pool of something that we want to access, but it's cheaper to just allocate a new one than it is to block waiting for the pool to be available. So we do a try enter and if we fail to enter the lock, then we just allocate a new one.

A

I could imagine similar situations with you know getting account.

A

A

U

U

Right, are you happy.

Q

U

A

B

Nobody else seems to make any faces.

D

I find it strange, but I really think that try is for things that are dynamic, not static, and we wanted to be static. We want at some point in the future when you know we are darknet right to even not worry about it and it's a strike and count forever. You have to like it. I would much rather be born again.

A

Then you're still gonna check it you're just gonna check it the results rather than calling it it's.

N

Basically, checking false versus tracking maker.

B

I mean probably most of the checks will be larger than their online, so that will oh you're.

K

Just gonna walk out making.

F

Sure how many people are actually.

H

Like that was mostly arguing the Decrypter types, as we expect essentially, then one can fit in the custom, tight or and who's the only one that we know that's ever actually derive the Dumars time.

B

We honestly throwing the nice thing with fidelia's that.

A

Means everyone cares needs if they're using some the other day. They have to know what the implementation is going to do, or else they need to surround every call with a try-catch.

H

That is clearly a blog I'm, a part of the person who wrote the channel reader writer. They didn't support it, but.

E

At that point, they're just making abstract like take the breaking change just make it abstract force people to react to them and.

B

What the difference is that episode will guarantee break everybody.

E

Do you have no evidence anyone outside of our accommodations.

D

Completely doesn't work, I wrote my code, it does work in and now I didn't even upgrade in my location, basically I loaded, my component in an hour that forcefully updated, diverges and now I'm talking.

H

Yeah otherwise I would work so to make the flower happy in the implementation in a virtual, zero or virtual negative one in a breath its contract, you upgrade your ref. You must implement this method.

B

You could do that too. I wasn't.

E

That your soft abstract idea so.

B

My third episode there was a customizable yeah.

H

But we can with because we have split graphing implementation. We can make something abstract much better.

D

Because it solves the problem of basically some libraries or seeing an update, one that doesn't even call this property and now the older version breaks, because the lottery reduces the load. That's.

N

Also, probably going to cause a bunch of work required for our reference assembly generator, make it no longer match the.

H

Attribute so we'll have to come up with whatever the soft marker is and again the soft marker in the CS. It makes it at home.

E

So if, if negative, one is a legitimate return value, meaning I don't support this, then why make it X crap in the rough you've defined a legitimate return value, so virtual, and so which means everyone has to chuck now. Yeah.

N

I think everyone has to check in either case. The question is: do you have a magic value which is negative one, or do you have a concrete I returned about your valid value person and people also people? So if we view the magic value, people are going to do it and they're nominally not going to check the docs to see Candice return negative one. Where do we do?

N

We have an API to check on which runtime version of Iranian currency do or net Corinne above you can correctly get the version number for that court to two and below it returns. You a crazy version number that you can make no sense about, but.

D

They never you never it never kind of conflicts can I write based on it said if I'm on that net five and call this property, otherwise don't yes, you can write that check and it's gonna be reliable, but you can't do it for two two or below it. That's fine, because this property is not going to work on 2.0 and four, we kind of have the tester.

D

The test is a my own data at five and above.

B

Apology is not really about that fire, but what you would really have to check is the version number of challenge itself, because challenge is out of that.

B

So you can totally up click the library running on their own on your neck or a virtual thing. But again, like you can check the version number of the channel library right I mean you really want to do that you can I I.

N

Think if we do the pool that the trial pattern, people are going to see that it could potentially return false they're going to check the docs in what case cannot return false we'll have a document did that this will only return false if someone's not implement. You know and then they'll handle it from there and.

C

N

The natural, if private value then do stuff otherwise other logic that they'll have to have either way. In the other case, you're going to have magically returning negative one they're not going to account for it in the first place, they're not going to see that there's a potential bug until they happen to remove. It simplifies.

D

The connection to the try pattern is that to me try parent is about like dynamic things during the program. I call try method, it can return. You know either true or false, depending on the inputs and state of the system. Well,.

N

In pursue, theoretically, someone can implement when it can return, but.

D

We don't know what I'm saying that nobody to.

N

Do frankly, should it's like it just gonna make the API card there to you. It's just like try get value on dictionary as long as you're using valid keys in the dictionary it will never return false. Theoretically, it could return false if you give it a bad food, but in the normal state of your program that won't return false, but for parsing api's as long as you're giving a good string if it won't back people-people program for the year, it's returning true case, and then they have exceptional behavior for the false case.

F

You see an application being written it, bring it bring it truly an optional API they be tested. Well,.

B

F

B

A

Diagnostic scenarios like, if you were writing a you, know a monitoring app. It would just not monitor it if it return false yeah.

A

I, don't know what the alternative would be.

F

I guess I am please.

A

Don't ask like.

F

If it returns false or true, can you make the decision in AB still function based on yes,.

D

So if I understand correctly the scenario that Mike- what is it's just additional diagnostic information, it makes these debugging easier. Oh the. If the count is not available, then his debugging would be harder. Well,.

F

That was my phone as well, but it looks like we already have something even more.

F

A

F

A

Is I I want to have a every ten seconds, I want to pull the length of my queue and I want to be notified. If it's getting too long, because that suggests something has gotten out of hand. Yeah.

D

I said debugging I've been doing Diagnostics. Are you seeing you know? Monitoring operation is crazy. Well,.

E

Then it for the scenario that you just mentioned, Steve I, would actually probably want to fail the application if the channel reader can't give me that information, because it means I, can't monitor the application successful anything.

S

A

Means you are unable to use the entire implementation that is otherwise fine. Just because that one thing you want to monitor, isn't available. So what are you gonna? Rewrite it from scratch? Know.

E

What I would do is I would block it in staging and wouldn't even let it go to production. I would tell my development team. You have to create something. That's monitoring this.

D

Thing: it's not monitoring the count, he's monitoring many many things from time to time. The monitoring says they are failed and then he analyzes the traces and because more information in traces, it's easier to analyze. What I.

F

Mean maybe I could see using this property as a way to spin up more workers. The readers not.

N

Right, yes, that was the that was. The reasoning behind the similar counts were exposed for the thread. Pooling TI's, which was people by default, can just use the thread pool and it will work, but in cases where there was a backlog end up growing too much, they want to have customized logic to say, let's distribute the work more or let's hold off on cueing, more work until such-and-such happen, and that's basically what this API allows. It allows you to say: if there's too much in the backlog, then we can fork and do some different logic.

N

Otherwise, we'll just continue working, as is in worst case. You get some more performance, it can be used for the tool which I.

S

Just say most of the purpose.

N

From the guys original pose.

B

I mean let's go better, so I think if our intent is, everyone generally just account I think getting a virtual count that throws and then having a way for people to know that they should override this. If things seems like the right approach and then it's just a function of of your cave of doing the abstract interrupt of the same Levite virtual intimidation, intrigue or are we doing you know a soft virtual thing or soft a perfect thing? If someone has some attribute analyzer or how they going to do that now,.

E

The call site would look nicer if you had a can count properly and account properly the through that way. If your code doesn't handle the and count you just get the exception, like there's no magic number like there's no try pattern that you always have to handle. It's like you get an exception. You don't have to handle it if we don't want to, if you don't handle, if it'll be on your application,.

N

But aren't most people going to end up using this for special logic or keeping their normal log they're, not necessarily going to say, use this logic or fail, in which case the eye pattern fits that bill know.

E

So I think some people literally just call the count property and if they care that there might be an implementation, it doesn't support it. Then they call candy health beforehand. So.

A

I'm looking at, we have three implementations of channels. We have one, that's bounded. We have one that's unbounded and for use with any number of producers and consumers, and then we have one that's for use with a single producer and a single consumer at a time, and the comment that is on the internal account property on the single consumer single producer.

A

One is this method is not safe to use concurrently with any other methods that members that may mutate this collection, which suggests to me that without a lot of due diligence investigating the implementation, and/or changing it, this will not able to this. One will not be able to return account.

A

So we would I, don't think we'd want to throw.

I

But if we, if there was a tester yeah yeah there would there would.

E

Have to be a test run and now.

A

I need to make two calls you.

E

Need to make originals.

A

Why is that better? It's.

E

Not better, it's just more usable is.

N

It the same number of steps as trying it count just with more guys. No.

E

Because try try get count like is going to return false in Elgin ends. If my application, if I don't want to make my application, have to even think about that. I literally just call the count property that's out under school.

N

Because if it's true you're going to you you're going to want to get the count, no.

E

Jacket now I have to write the if statement like you're now forcing me the caller to writing that statement. If, if I didn't know the house, we have statement well, what do I do, then? Don't.

A

Do this that you would do if you didn't want to check them to, can we can't count yeah.

N

If you want a ballot count, you have to check.

N

So you're either you're either not going to use it, in which case you're not going to check house count and you're never going to get a count or you're going to use it, which case you have to check cast count. You have to.

E

Gain don't like the scenario that I mentioned a few minutes ago is like I'm writing the simplest possible monitoring application. I expect that the implementation I'm using your work I call the count property because there's also a can count tester if the count property throws. It means something is wrong inside my logic and I need to go trusted yeah we're a ghost apart.

N

That's also, if not I, get count right.

E

Yes, but no, no, which is normal.

N

Exception handling laundering for like everything, but.

E

Now, I have to be aware of the fact that something might fail, as opposed to just having it fail naturally, and letting the rest of my monitoring system to type in my application has failed. I also.

D

Think that the comment that Steve was talking about in the implementation, I I might be wrong, but I think we should just implement this come property even on unbounded channel. It may be, like you know, there's an internal thing that you know nobody wanted to spend in the past time to make sure that it's.

A

Suggesting you suggesting to figure out a way to fix it or ditch this implementation.

A

I'm not sure it can be fixed right now, because it's a super low, lock, there's zero interlocked operations, zero locks in this special case, where there's only a single producer, single consumer count, requires more interaction with the entire list, not just the head.

D

A

It's not about getting account it's about crashing.

A

Because you're you've got if you have multiple threads that are interacting on the same state: you're gonna with no locks no interlocks, and it's not safe. To do that, you can get a reference exceptions, you can get a visa or what I like you can there are lots of things that can go wrong.

D

A

Not reading it, it's walking a list, it.

E

A

Store account there's no, that source account.

E

The jeweler's for the anticipable.

B

I mean the difference between tri-gate count and has count count. Is that I think to the has count is kind of clear. There is aesthetic property of the object. Well, you're not supposed to check this everything operationally like it didn't very end of the of the obviously and.

D

Secondly, I I go back to like the rally. You know we keep able to make try. He was here. The try guideline say we have some api's, that throw exceptions and performance of the exception is unacceptable, is an alternate. It's basically a performance of.

B

Which would be here as well and you can you can still access, town and catch. The exceptional I will.

D

Didn't I, don't think we are saying: hey, don't worry, we will design it with an exception, or should we have a performance problem? Let's add a try. That's not what you can see. Okay, the guidance we have on try is if.

H

You have a try at an on track. That's enough! Yeah I! Don't.

Q

Know you see I didn't so actually I'll just do.

H

If you have a try, you need an on try as well and mannequin.

A

City, we would never add an on tried, EQ or NQ method to these types.

H

I guess yes, we'll be concurrent ones, try being something different if.

D

You have a target I wish, but not yeah, but you know still for DQ. You know you could imagine an ID ID throws. Forgive you if there is nothing in the pew.

I

D

So he added a try burden. So to me, it's consistent that each of you is. It doesn't follow that. Okay, you have an Android version, aren't of the guideline palette. East II guideline is used or the try pattern is used or what it was designed for. This one I would say it's not. We can change guidelines, we can keep adjusting them I'm. Just saying we have a guideline for exactly this thing. It's a you know, new support that we want to add without breaking changes. We already have a paranoid and it's the desperate.

D

Do it not try and it has the advantage I find that there's some complexities here with you know the single purpose of the channels, but.

B

What is the complexity? I mean you could just return false for those guys, I mean it becomes an optional feature. You can decide to not implement. Even if you wear the feature right, the.

N

Tester do wear pattern is just the try pattern with.

D

Two API, rather than well no, but also with the implication that it doesn't change over the lifetime of its bit. It's a property of a time is.

F

Not the problem I.

D

Know it's not the state of the instance or equals. What.

B

Is the state of the instance.

N

Right we, we have other places in the framework where the eye pattern doesn't necessarily change, whether it returns true or false.

N

The floating-point parsing api's, the way after usually yeah.

F

D

N

Triple E combined ready to do it. No.

D

No parsing API suicide because of performance. Yes,.

N

But but but for floating-point they can never fail. Oh yeah, our first Q double.

M

Dot Triforce, Q and believe that returns the super.

F

M

That's a fail right.

N

But that's not an exception. It will never throw an exception. Well,.

H

The highest method, rule right, yeah, double dot, parse will wear, I want a skate deposit and.

B

The try part shouldn't throw for nothin put the simple point of if yeah yeah the thing in this case I'd to me. The reason why I think the test that will make sense is because I like using our earlier I think you don't want to care about it. You just you just directly all count and it blows up right and then I think in device scenario. The idea would be something very high up in the stack that actually takes a channel meet or produces a channel meter or a channel writer. That's the part of it.

B

You may check. Do you have a count? No okay! You blow up earlier and.

D

Also in the five ten years, like literally, it would be no reason for you to check this.

Q

D

Basically, just know which channel used- and you just called the matter, I.

B

Mean it's a bit like scream I, like you know, can read, can write, can see right depending on what scream you get. You have some combination thereof right and most code never checks it well.

D

But if I don't.

B

D

Exactly I think it's analogous if I use you know, memory string, for example, I already know that I can see. Why.

B

D

Ever check you never check on memory, small.

B

For the lady see it's more for the receiver right, so let's say you're right, X document dot say, but something most people don't check. They just directly try to write to the stream and if it blows up or than the Koala gets an exception that says you passed me something readable, but it anything right of all it.

E

B

E

What we almost want is something in the language that says: here's a try do something method, I'm just going to call do something, please language or something else, wrap that try call and then throw an exception if it didn't succeed, because I don't know how to handle failure online well,.

B

Really have that right, I mean you just we.

M

Just call it Colin implementation hopeful for you, I might like that. I mean more generally.

N

I think there's also a difference between the stream API is on this. One stream api's, you have can read, can see and you end up performing a function. Call that may not return a value such as C, whereas in this case you are clearly trying and then getting it out and there's a clear difference between those two lecture.

B

Understand because I mean well.

N

In one year, you're you're you're doing a check and then you're receiving a value and then the other one. You are doing a check and then performing if performing an action. Well,.

B

I mean: can we do returns, cool and then I call read that builds a buffer right.

N

Right and Seto, for example- oh, so it's a it's a difference of you generally want to have an output on the fall scape right. No.

B

They mean that if you want X document, don't save you cannot operate until you get a screen that can write. There is no there's no L skate. The odds cases write different code. Give me a difference between. There is no alternative opportunity.

N

But in this case the rate, the alternative is the trends.

B

Of what you do every no, it depends of what you do, but if your whole features monitoring your fundamentally cannot operate. If you can't observe this thing, so you would basically, if you, if you wrap a channel reader or do something to that extent, then you would forward it to your to your consumer and say you passed me the wrong channel either channel writer, because I can operate right, but.

N

Is there ever a case where you would want to check, get a true and they're not immediately.

B

Yeah, because, like imagine you were writing a monitoring channel via the channel writer, where you basically construct the channel reader that just wraps another one, we would just fail early in the implementation and say you must pass me something that is, it has account. I will not work if you could and then later down like when you actually like, go, read or write, and you do the monitoring. That's when you actually call come write your.

V

Concern with the can count count method now, people's perceive can't count to be accurate, so here character, returns true and count to the equity investor.

E

V

Accurate 1440 lecture from Bible. What, if is that going too far, a can't count? This property on this instance returns true. That means squarely accurate.

V

H

Can't be because you read the count and if someone called that or some other consumer Oh EQ like count is that wrong. I know I, think if you have it the.

V

H

At any point in time, and unless you were the sole producer and soul consumer, what you could be you.

V

Don't know what.

H

Is the acceptance.

V

On property look and that's what the can count: guarantees I.

B

Think, there's a difference between the test that doer is not about what this operation succeed. It's about! Do a half this feature right. Thank you think. If you have you know and right on the stream. Well, if you discus for and if I divide the divides that you will fail right but like like.

Q

The father basically.

E

Been calling right, moral of sporadic so yeah, that's really what I mean yeah.

H

Henry and on screen says: if this is false, we will not support.

B

Do we want to call it cast and collar if we want to call it has come because and she can't we got away really what I call it can code and.

E

B

Ask if you got exactly.

M

Commando method itself.

C

E

Killed, rather than use a property then can account.

N

Is gate count supported.

E

Normally we use can, if you look at strength as positive.

D

Not making property because to indicate that it's performance so.

E

I would make it about.

D

N

The guidelines generally say that there property owners should be simple operations and not happy.

E

Pretty hot stuff, alright.

B

N

B

I mean there's, nothing, I can do it like Network is Network it's and I cry in a corner. That's the best. I can do.

B

All right Chris, if you remember your ID, probably not right, that's what I feared.

R

Did you file the issue? You did right.

Q

H

I

B

Why the is it not in the a peer reviewers this.

V

L

Suggestion Oh somebody didn't know, follow the process.

E

Is it ready full of.

B

B

This means you're, always beer.

L

This is 405, you.

V

Look, no, the API is ready for review and you'll set the master.

B

That it makes a difference, Oh God. This is only a few again Oh mo.

N

Have you up in your tools, so it's tracking down that wrong. Now,.

B

I mean it's a config file change, though yeah let's say: I have.

O

O

I'm, making a signal like you started the data when your hashtag Yahoo and anything goes how many entries are use unless the original iconic doesn't happen for the.

O

That we decided that we were just.

O

And the the offer, the other specs later like a asperities, a different wish, Marshall I think basically suffice and one sequence as the stupidest.

O

O

This basically asked the ask the owner knows: is former I'll be like how not sure and kind of so basically like all of the matters I'm having here is I'm sure I did. The color of this picture is like sign hush and we on DSi and DSi and then versus enough like helper knuckles. Here.

D

Make sure I understand you? Basically just we have those methods, you just added overloads that take this new enum value and then maybe some helpers there, except for their protector before, because we already have.

O

And one month, I had a horse allotment of a lot get much signature spice at my ready, Thomas I mean yesterday with evolution, the output, the signature or useful, otherwise,.

O

I said these four dozen young.

D

For the put and based api's.

D

The rights to a destination and can it be, should it be operation, especially I, think anomaly, there's no way to continue is the distinction is not large enough after we start for about.

L

E

F

O

E

Teams in its entirety or fails so there's no operation. We've done us. The.

H

Amount of data written before I Triple, E d13 60, to be on any instance. It will always output the RFC 30 to 79. You can all sign over and over and over and over and over again, and it will fluctuate essentially plus or minus four bytes.

H

Oh that's why that one is that's. Why the answer? It's not get signature, sighs it's getting max as there's as part of the algorithm. It has to be like the native layers generates. A random number does math and it comes back and says this with you essentially will just never call it unless you give us the fingers bigger than the maximum size, otherwise OpenSSL will happily do but I don't.

D

Answer so say the size of this signature or its size of this. It cannot be size of the sine theta, because this is what about this. Api try sign date are called. It takes the destination and I estimate the size of the that's. What the get back see that doesn't think the signature size is always dependent.

O

On the size of the.

H

Signature length does not vary based on the signature, is always a fixed size plus or minus something. So.

B

I just do PSA and ECDSA like they seem to be identical, except that one is called signature. The other one is called sign likes either create signature versus sign hash. Can you explain the difference between those two.

H

O

One they called.

H

It create signature on DSA and sign hash on RSA and no one uses VSA. So when we made DCPS a we made, it match RSA.

H

By data inside hash, instead of signed data.

D

And I also assumed that the casing of those acronyms we.

G

Grandfather, oh yes, a new city si their.

E

Existing types yeah, the well speaking of casing, like I Triple E, should be all uppercase both well.

N

In the five of the other places, we've used it it's lowercase and we only have an uppercase in system map. I could be there Oh in all the other, like networking and other after I lowered.

H

Its the I Triple E, oh, that P 1363 eh.

V

Sometimes you see it written with an e on the end, and sometimes you don't. Let me assume that that ability is enabled for all these api's. It looks like I would say yes and therefore nothing changes. A binary should come.

C

Here, our stream.

V

Can't be now, and the other thing is you: have this type of called dsdm use your format? Should you be the formatter, because I don't find this type anywhere? It's listed here at the top. Isn't you know.

K

What's being added in an author, I also consider.

K

That is this. One is made of pipe.

K

In your classes, but we haven't been maintaining those.

V

Okay, going back to the dip like really try and verify signature versus verified data sign the parameter is the consistent, but you see I. We use our BD RGB hash.

H

They're consistent to the classes that they belong, they should not be consistent with each other, even.

V

Though even do a span, API has it called spat hash, but the binary key is called RGB.

E

H

I, don't remember why RG is it's rendering venerated lights? It's.

E

Randomly generated by that argument array, no, so our RGB in the context of this basically means that a hash function is a pseudo-random function, which means that the output of a hash function is basically random bytes, so RGB is Rena.

H

West we're in office cuz, the RG will always use the prefix for anything that.

E

H

E

H

For range yeah, so sorry and we generate.

E

Its range of its range of mites, oh we've, no yeah, we've always used it as the output from a DRG, B or D or D.

E

V

Is this a turning point for the API here, Hispanic guys, have you again the RGB.

V

Glad I'd give you the bad name.

H

We're overloading create signature, the rules for oh yeah.

V

H

Do not vary the name between overlook, but.

B

That's what I call it data because existent one is go there.

U

The hospice be.

H

And with RSA signature or matter what we call I didn't we keep the RSA all caps, because the class on it.

E

Or you say yes, I triple he's also trigger, which is why it could be on capital.

O

Thank the only change me.

E

Or casing, that's the only I would.

M

Obligate and we can not done it elsewhere, except for the math, a project.

B

I would try to think the we'll be having a crypt already like if crypto already has it in lower utility.

E

We don't like ten or so the only place where we already have. It is in well.

N

We've got it in system numerix. As I triple e 7 5 4 helpers, lower case system net primitives for sockets. That's I could lead 12, 8 or 4. So.

E

I, don't know where I.

V

Don't see I'm truckin api's I've done ooo there. Ok, this is seminar sockets and just man. Videos I have to believe lower case. That was probably a mistake.

V

Lots of old data that you please stand for by force also.

B

I mean given the diseases whenever, when the price were beautiful, yeah I, honestly, don't.

K

N

I three just couldn't see sure to allow superscription that we can do. I IE q is that one.

L

Small capsule you.

B

Know literally the fun thing of like browser windows: they keep popping up in all over existence all right. So let's do this. Let me yeah all right. So then, what's really keep the abscesses here, then no cab. If the caps is long after.

G

Get the capsule kept as it's listed here other.

E

G

D

Say it reads like yes,.

E

The thing is like I: when I introduced the utf8 helper type like I, put lowercase: u lowercase TF, even though we used an uppercase everywhere else, because I'm like it's, not a trademark, it's just an acronym in our case in guidelines, say after and surely we have the first thing capitalized unless it's a proper name, but.

Q

Does that not capture your feet into this country, yeah.

D

Recently, I was replying to some email, not email, a github question from a customer. Why? We have complexity in the casing, guidelines? Why? Don't? We just always use Pascal pacing I saw I said it's because system area would look like a namespace for manipulating and goona Jupiter. Yes, I, like.

B

D

Others responds to that by the way. What I'd like Arthur's response to, though I didn't see more in his.

B

Response was thank God to be aboard that confused.

E

Just wait until we actually have to add something like iPhone to an API name.

B

The guidance said avoid trademark.

L

O

L

O

Q

Oh more obvious and look at it, then it's like.

O

B

Let's face it: if I have to use the API we'll hit f1, there will be some code sample of a copy and paste there's Burger.

E

King sat into mice code.

B

And then move on.

E

B

That or we'll be on Stack Overflow, like I. Don't care like I'm not going to like go through a complete specification, but I like let's be real. Do.

V

We expect more names or things to be added to this type. So in that.

K

V

That our parameter list just keep going but I know your hash algorithm at the enum horse-drawn. No.

E

They, wouldn't you, wouldn't, have additional parameters, you might have different entries and uses.

V

Oh yeah, I'm ugly, ugly.

E

New types of enums, they would rob at that point. What we would probably have is something that's not DSA, so we would probably have a different type: a different class apart, okay, Oh.

V

Presumably, like later I'm looking like parameters right, you have data officer, town and then hash algorithm, and now he added you need a format. Yeah.

B

V

Expected to add more or ap is more grander, Susie I, don't.

E

Think so because again at that point, we're really talking about different algorithms, yeah.

P

I mean the way crypto works is that it's very fixed in terms of inputs and outputs and I. Have you with those? Then you pull out the different a little anyway.

B

V

He added this, but.

B

This is like we like. What we added here is the weather get this now.

B

This hoop is impossible to see, but basically like, if you, if you squint it's basically just it's just a different way to do the exact same operation right. It just is like, even if you look at how these things are constructed very often like, especially with span and non span yeah. All you do. Is you just change the way the inputs and outputs are being passed in, but you still have the same time. Well then put it on for three yeah.

E

Then probably the closest analogy to what you were asking for awesome is: when we added gzm support it's a different mode of operation or algorithms like AES. The thing is once you once you start talking about stuff, tucked out it's so significantly different than we actually introduced a brand new type board, because some of the other parameters simply made no more sense of that would be default.

V

They exist game over those like when the Monday takes by today by today, hash algorithm.

E

No because they're the existing, the existing API is already had appropriately chosen default. They just don't work yeah.

E

So there's DSA dot, sign data already yeah.

E

So chooses an appropriate yeah.

V

Like what I mean is since we're adding a new parameter, there was a DFA format. Yeah. Do we and.

D

I can ask presentation, you see leadership, the utf8 shrinks and in Sanders this reminded me and we need the URI now. What is it called tightly they're escaping that we don't support 2064? Is they thought.

V

I was in perhaps no we that I mean that was all my plate because Turkey okay, just like we have you basically for API yeah.

E

L

You're, not you're tracking up on the other icon and why I know there was an issue, for there is no issue.

L

You see, that's not distorted that difficult yeah. All.

G

V

This one some reason we remind unless you have something else like where you need more cost data context for this. It just transforms right.

Q

Is that ready, yeah.

N

Pushpa offline.

B

Is that the one who 500 don't.

N

B

N

Those are existing, so basically there is on x86 there's two encoding. There is the legacy encoding, which is basically all machines prior to 2011 and then there's the VEX encoding, which is for a VX and above these two have different semantics. When you have to consider what memory loads can be folded into the instructions, so you get one instruction rather than two on older hardware.

N

Those instructions were required to be aligned, and so we can only fold a load aligned into, for example, and add all, whereas on Beck's and above they are unlined they don't they don't actually check the alignment anymore, and so we can only fold unlined loads for the most part. People writing api's will either always do on the line or they will.

N

They will pin the array force their data to be aligned and then operate on it as if it was aligned in for the for the entire data length. So the proposal is we expose this API, which in debug mode will validate alignment regardless of whether you're on a VX or legacy, but during release mode it will always be folded to be correct load. So it will be an unlined, no check load on a DX hardware and in a line load so.

B

It's typically, what does it do? It throws an exception. All right, right.

D

It will throw an access violation exception, but underlying loads. They.

N

Not work they're, just slow, so unaligned loads are on modern part where there, as fast as aligned loads provided you're, not crossing a cash boundary, which is why many people pin their data and then ensure the data is aligned regardless. So this is just a convenience, API to say: I done the appropriate checking to make sure it's aligned during debug mode you'll get the additional validation that is impacted line, but we will always, regardless of whether you're on old hardware or new hardware, omit the efficient encoding. But if you own.

D

Them so if you want a new hardware and it crosses the cache, you get a purpose, it's it's still more than sanguine is always just and then on all the hardware we do we ever failed, or it's always also in just the perfect. It.

N

Depends so if you're doing, if you're doing load aligned, it will fail if you're doing load on the line, then there'll be a perfect and it's even more signal and we.

D

Don't have X please when users don't say guys, they don't explicitly say whether it's slow, the line or the.

E

Existing API is there or yeah.

N

The existing api's are so few school that actually they're right there at the top. The existing you guys are load on line which is just moved back to 120 and load with a minor. Yes,.

D

N

The top API would do load aligned on old hardware and low to unlined on new hardware, but on new hardware in debug mode. It would also validate that your isn't a lot so.

E

We do actually call it load aligned in a few places today, especially in code paths like what Tanner mentioned earlier, we're like a lot of our ASCII process and stuff. We pin, we calculate the alignment manually and then inside of a tight loop, we call load aligned. Are you suggesting the code like that would be changed to use the new API right because it'll properly fold into animals.

N

Properly fold, it will be more efficient and allows you to write one code path that handles both cases rather than having to write two code paths with it. For example, explicit debugger.

E

Why would I use this instead of unsafe to operate on aligned of dr. 128 afloat, because.

N

Read on the lines won't be folded on legacy for work, but.

M

This will be better than.

N

M

Well, people what legacy is.

E

N

For example, the MacBook I have supports AVX, but not 80, X 2 and so, for example, from her 4 or for some combinations. Also for a ot scenario, where you have to compile for legacy non vex encoding by default. That's always going to be aligned, checking aligned folding in that scenario, and so we we end up with slightly less so.

E

I see I, see utility for the name. Unsafe, probably isn't really sitting there with, which is normally unsafe, means memory unsafe, because you leave a point. Okay, look quarters, yeah,.

N

It's pointer yeah in this case it's memory unsafe, because, because.

E

You could maybe right.

N

Meet me because if you're only testing on new hardware and you happen, have a line data and the GC didn't move it, then your test can theoretically pass, but in practice that's likely not that so.

E

You said on new hardware, there's still forces and will end access and debug yeah.

N

Well, in yeah in debug, it will emit a low blind instruction and not fold, but in release mode. It will appropriately and.

V

The dog mode user, zapped Andy Belarus pay.

N

E

Glover's, certainly sorry sorry say that again.

V

This behavior of debugging is happening. Yeah.

M

V

N

E

How do we know what the user happens, compiler optimizations.

N

Neighbors are not there because.

B

Digit knows what they you want to have up to my sport or not, and it's based over yeah. So if.

N

Your compound on tier 0, it will do or if you compile, for your one or aggressive optimization.

N

We'll just do online check, not in hardware, so.

E

If you're running under a debugger, then it will do the alignment reckon that.

N

Depends here yeah.

E

N

On how is that whether it was compiled? What whether or not you have the suppressed auto optimization is tracked and I'm.

E

The diva babe, if this, would you kid yeah? Well, because no normally, when we talk about debug versus release, we talk about the generator, the il not about what the JIT is code running normally yeah.

N

I guess this is optimizations are enabled.

E

Winter often reminded me, winter optimizations enabled versus not enabled they're.

N

Disabled on your zero under debug mode, if you suppress yet optimizations via the environment variable or the checkbox and they're enabled for tijuana aggressive, optimization release mode. This.

E

Means, presumably our unit tests actually would run for an afternoon. It.

N

E

Could you introduce that the ability to run them under and are not optimized good.

M

N

So we for unit tests, we have the ability to run with each of the environment variables for enable and disable, and then we always run in I think check mode. Okay, isn't.

E

The reason I'm asking is I'm trying to figure out because we're I'm, probably the biggest consumer of these API, is in Corsi avoir, so I'm trying to figure out like how would this affect the code that I write and we.

D

Considered I mean 128-bit aligned.

N

So there's going to be an API, that's already been approved, it's just not been implemented, yet that allows you to allocate a 3-pin array that can be aligned to whatever about you want. The GC will never move it as long as it's alive, and you can do that, but individual instance members on the heat we've got no way to currently a lot into 16 byte batteries and.

D

Those special arrays that you mentioned a they are still DC arrays right. There.

N

Regular arrays, so just it's just a we're, not nothing more yeah.

N

No, it's a new pinned object. Well,.

B

There's a new HEC, AP area located away. You can specify you wanted and.

D

Would you would, would you rather use those, and you can you know, like I assume it would be.

N

Better to use this new heat, there's no there's not necessarily any guarantee that the input data for a given algorithm comes from one of those yeah.

E

Also, like the algorithms that I'm thinking of all to expands, which means they've, they don't have.

N

E

N

Trivial to write an algorithm, which pinzhi array checks, the alignment windows at most, one that most to unlined reads or writes to deal with rail. It with.

N

Cradling and proceeding elements that are on the line and if.

D

Those new arrays will I be able to check whether, like, if I, don't see, if a spanner or an array to my algorithm and I check on which, which I don't think, there's a name of that. I.

N

D

N

Cheap that that'd be more expensive than just pinning in checking I.

E

Mean if you, if you already, if you already have a span coming in like you, you have a pointer to the data like the data is going to exist in memory somewhere. You don't get to control her. Those. The only thing you can do is react to it and at that point, like header said, just didn't, go yeah.

N

Because it because it's likely cheaper to pin and check the alignment and it is to have the GC check which well.

D

I would imagine we automate those arrays and we put it. We would put it somewhere in there under cover this space. There.

E

Is a yeah, but we we do have. We do have space but checking the object. Header already implies that you've pinned the object and if you pin the object already, you do have to finish up the offer cutter. It is mandatory yeah, the GC can't track reps before the object method. Data I mean I.

E

Think Tanner had a proposal for like how to fold it into a a like two parameter structure: there's something to do that fact, but in a two part benchmark format in it, but yeah actually you'll see wherever we chuck the bits in the object header. It always involves me. It's just a consequence of how the GC works and it also bolt interlock, oh yeah, so I guess the other name in suggestion here would be. Should this be load a lie, detector, I'm, saying because it does require alignment, at least in debug, I. Think.

B

That's the thing others understand. So what is the intent? So the intent is I call this API when I know the data is aligned is right well, as humans online, its alignment. If it's not aligned, then in debug I will crash and then release this purple stuff. Then.

N

In release on new hardware it will suck and on earth will be less than ideal and on old hardware you will get the process violation. I see.

B

In the debug view, make sure that you always get an AV, no matter.

E

What the hardware right, basically I'm I, would have written her load, a line vector 128, but now I'm writing something else in place of load, a line vector 128 that.

N

Way you get the efficient code on both new, which.

E

Makes me think that I probably still want the world of lines somewhere in the method? Name, because the method is not expected to see that if the data is misaligned, you.

J

Ship together, again, you've already shipped in court. Okay, yes,.

N

The other ones are existing. The unsafe I'm.

C

N

It load line, factor 128, unsafe yeah.

D

So you're so you're, okay with the line of an alien but yeah. So we have these met. I think we have a convention, we call the methods, throwing we kind of have a very ended throw so it doesn't throw because for you and say, basically means it. We throw great well.

N

So it's no unsafe means that the that it doesn't strictly always do the alignment checking it's unsafe because it doesn't always.

M

E

Alignment checking does.

M

A lot a typical CLR exception, it's actually a process or maybe which the CLR captures and converts to an access violation.

P

But you basically at that point: your process is kind.

B

E

Typical Nora, no.

V

We exposed the encoding legacy verses backs in the API surface, right I being lied, no.

N

I, don't I, don't think that that's important. These are already low-level api's and people who are people who are using them are expected to already be familiar with how they work and look.

E

This would actually optimize some of the code we've written internally if we could take advantage of this I get.

V

That, but, like the whole motivation for these, this API is that our vex versus I just want reading legacy they're different, holding in or in semantics right. Is that something you just talk to it? Yeah.

N

That's something that we'll probably don't.

V

Want expose that no.

N

I, don't think it's if.

E

If I as the caller, these api's had to say if legacy encoding this otherwise this like that, introduces more opportunities for failure into my own code, because it's now forcing me to duplicate, effective work.

N

And in the primary purpose of this, API is, if you want to support hardware for like the benchmarks games, that we that we've written stuff where, in the past or modern hardware, you end up having to write basically the same code twice where the only difference is the line versus load on line yeah and it's less than ideal. And so there's API solved that problem. Yeah.

B

So one question I have is so: what is the behavior of load vector 128? It basically means I, don't care about in lime and just make it work. No.

N

So load vector 128 is always an unlined load, and so it will always succeed, provided you that the memory is beatable or writable. Look on new hardware. That means it can you fold it into the instruction, but on old hardware it can't be folded so, for example, in a ot scenarios in cross-gen ready to run in all those scenarios. If you use on the line for the better for the better tier one through point, you end up hurting ready to run scenario.

E

So so, if I can also help simplify what Tanner had said, imagine where you have I want to read this memory address and put it into this X mm register. Read this other memory address, put it into X mm, so I now have two xmm registers now I want to add those two things together into another X mm register.

E

What Tanner had mentioned with folding is I can do the first read read: address 1 into X mem register 1 and, at the same time read from memory, address 2 and add it directly into X mm 1, for instance, without also blowing the accident to register. So it's it also helps make them register allocator more. If.

N

You scroll down just a little bit. I've got this well.

B

The reason I'm asking is because I'm just still struggling with the naming so basically what's the difference in load between load, vector, 128 and load, aligned, vector, 128 mode.

N

Aligned always does alignment checking and so on old hardware. It can be folded into the instruction and on new hardware it can't be, but it always does alignment checking yeah.

M

There are actually two different instructions which one is then the one I should be using: no no right to file for.

N

A new hardware, which is the purpose of busy-busy.

B

Was saying if old hardware load vector, if new hardware at low the lion vectors no.

Q

B

N

B

N

Be aligned because aligned memory is faster to access, but on new hardware you have to use unlined and on old hardware, you have to use the aligned instruction. This API provides a I'm I'm using the aligned, because that's the fast way to do it, but I want to do the efficient encoding, regardless of what hard work well.

D

Yeah the reason why I think emo and a bit confused may be other people in the room. When you say aligned has it has potentially two different meanings: whether the memory is aligned or whether the instruction that you're calling is aligned with it.

G

Sometimes it's.

D

Not super clear from the context so.

E

We I had mentioned earlier a lot of the code that I've written to make things like asking utf-8. Processing more performance would benefit from stuff like this, because we do aligned only reads right, but on newer hardware, which is primarily what we test on those aligned reads, won't hold into the next instruction, as Tanner had mentioned, which means we're actually emitting bigger Cochin. So.

V

Why do you align.

E

Rather than unlikely because I know I, the data is aligned and I want the access to fail if the data's, not because it means there's a bug in my code. So.

V

B

Use the new one I would use any one. So is it I'm afraid to say that, logically, the new operation is load aligned, vector 1:28, except that we used a better encoding that we know will not work on older hardware.

N

E

It does about sufficient thing, regardless of what generation of hardware you're on and.

N

The expectation is that the data is aligned and we will do validation of that in unoptimized code. Okay, so.

B

That sounds to me that the new API is neither the equivalent of load vector 128, nor the children of load and a lot of load and line vector, 1 right.

N

It says it's a hybrid of both in a sense, depending on what hardware you running on. However,.

E

Because it might a V if the data is unaligned, that's why I had proposed adding aligned to it so.

B

That's what I'm saying I mean my original purpose would have been unsafe, loaded line, vector 128, because that's kind of the closest we have in terms of the naming convention, but that kind of implies that it's basically lower line vector 128 except you, don't do the check. So you may go up so so.

N

I think long line factor 128 unsafe. You know Levi's requirements make it easy to find it would all we.

E

Usually don't use on savers and stuff is really on.

B

E

B

E

Stuff, we do because we want to keep the method groupings together. I think um I think you have other vector or api's aside from the intrinsics that do that you know, I, wouldn't want it up to something I'd.

N

Say because you want the method groupings together. Well,.

E

There are good team intelligence because that's until they send people away so doctor 128 or create scalar and vector 128 or create scale or unsafe, like we for a lot of the vector stuff, we actually already have unsafe as a suffix I.

B

Think we should be shoot honestly, think about the guy. I would I think we should generally strive from making it to prefix, because it is probably the first thing you should be reading. Well,.

N

B

N

Are already unsafe because they take pointers, it's hard sure like.

B

All I'm saying is that is the gentleman pad on the whole, but the whole point of the unsafe keyword or the sorta, not the key word yeah. The naming convention is so that it pops as soon as you make it a suffix. It pops a lot less because it's like song and they, especially when the API name already, is like twelve compound words together and it's like. Why are you doing a load operation? Oh yeah,.

P

Then action might IV, but that's something it's all the weight you know about.

E

I, don't feel strongly about including unsafe anywhere in the name here, because you already had to write it on a safe block.

N

To differentiate yeah I think.

B

In this case, it's like it's fine, regardless I'm, just saying like as a naming convention. Even let me use unsafe to indicate okay, this API.

M

Would be something.

B

M

B

N

I think in general, I.

B

N

D

B

That easy! Because what.

D

Are useful to name the parameters, they're aligned or underlined the address basically make it here. What the aligner underlined refers to, whether it's to the interaction that you cannot be calling on to the expectation of man even.

N

Using mister going to understand it's the address.

E

Because, in in the.net framework, like we don't give you the ability to control which instruction, what which memory boundary your instruction happens to be on, so we would never like, because we don't give you that ability in the first place. There should be no expectation that anyone wouldn't be doing that.

E

Look if, if you, if you needed that type of granularity, you would literally drop down the wall, something you can be like and so far so.

B

A condom: what's the name that we would go with no.

V

The line that come on 28 and say one question I have is: is there a possible for a feature? We I don't have to worry, think about alignment or a line or a line, but III thought that this API would let mean obviously, that I don't care what I? Unlike you, if.

E

If you don't care about it, just always call unwind and day which.

C

Is load back to 128 I see this.

V

Is more about obfuscating yeah.

N

This is your your codifying for the absolute best performance you're, guaranteeing your data's aligned and you want the operations to correctly hold whether you're on ready to run or legacy hardware or new hardware is.

V

Existing Lord, unlined, October, 2011 38, if you are, is that optimal for unlined with are smarter. Well.

N

So that always does an unlined, but on ready to run and on old hardware, it will be two instructions rather than one, because it's a limitation of the hardware and we're not going to go PAP CPUs room for.

N

E

V

All right so then, there's.

B

V

About there, what other api's we want to expose this double-stuffed, abs and or.

E

If you explicate you, the load align the vector 256 right, yeah.

N

There was there iin, a question of like: should we for parody, also expose the equivalent on ATX yeah you trying to do word searching. Oh.

E

I'm trying it doesn't necessarily.

N

Make sense to expose on a B X, because a B X is.

E

N

New hardware does.

E

Avx even have a load, a line, defect, arrow 256, okay,.

N

It's still does, ah but it doesn't necessarily make sense there, because AVX always supports the online. So you could trivially just do debugger cert. If you want to validate alignment, no that's fair um and what.

V

Would the story API.

N

These store ones weren't included because those stores don't get folded, except for any rare couple. Cases like extract so but I think extract always does online checking on I might be mission, but if there's a difference then yes, it would be desirable to also expose the same or.

V

That's something that you like before the future, if yeah.

E

I was going to.

E

Random question better: would there be a load vector 1:28 of you? It.

N

ah No because none of the hardware indicate has work on doing.

M

The only time they wouldn't have them in well, there we wouldn't be able to.

N

Yeah, there's no hardware for it. It's they either operate off. They operate on the ten primitive.

E

No I mean we would fix it up to the intern for long under the covers, but now we have a question of like x64 version. Yes, yeah.

V

Right I had a follow-up question for the biggest asset. If you can go back to that one, it's a good thing given.

L

They need a Edition number or two six, eight four or eight four.

K

V

So you have the API that they expand and return and out by certain, but the out parameter is not the last one very.

L

V

Five six ply signed data, oh yeah, because it's no big business as well yeah, so the ocean store still.

L

V

Be the last one for me: yeah.

E

No, that's a good couch.

V

There's a few are four of their for eight ideas that, like that you understand what about.

E

The hole over from calm days.

E

And when remember calm, you remember calm.

D

I, remember not being able to understand anything.

E

Do we actually have an H, Lazar type in the framelit? No.

N

Because struck rappers are not the same and primitive values.

E

Make a primitive age presenting.

B

N

Resulted in dragoon idol.

B

World is not a better place.

B

Well, there is B accomplice of it only.

L

You this is not annoying at all. Just you know, just asking the question: yeah, you get a cue in the combat said over.

L

Like, where is you know, yeah.

B

L

G

Also, really easy to edit.

L

Think I got them.

E

Really it's get fixer or you like right-click on it and say rearranged parameters, and it also picks up all of the call sites and all of the samples that you ever any now. Yes,.

V

B

Only it takes an iterative that support that you could embed in JavaScript right, I.

E

Mean this this entire, you know, weekly two-hour session is basically us running our guidelines through the Mechanical Turk. That is us.

B

I mean it's automated right, but you're, saying everybody who is in this meeting should change the title to a Senior, Mechanical Turk. Yes, that's cool and I'm, the I'm, the because I'm.

D

Exercising the mechanical mechanical, sorry, you are.

B

You're the what I'm the mechanical mechanical, because.

D

B

Having to type it in.

G

Principle, Mechanical Turk will be a principal to my back. That's it! Yes, that's how that works. All.

B

E

B

E

Principled mechanical.

B

All right, I think we have enough jokes for today and I. Think I will conclude this meeting and looks like we were still streaming to YouTube, even though, even though the network keeps coming in and out so that's good all right. Bye-Bye.